Improving FAQfinder's Performance: Setting Parameters by Genetic Programming
Edwin Ceeper · 1996
As the quantity of information available on the Internet continues to increase, so do the attempts to harness that information. One such attempt is the University of Chicago’s FAQfiuder project; given a natural language English question, the FAQfmder system attempts first to locate the most relevant list of Frequently Asked Questions (FAQs), and then to find the question on that fist most similar to the one asked by the user. Since FAQs nnmher in the thousands sad each one represents a concentration of information in a specific subject area, the FAQfinder system offers a means of accessing useful infonnetion in a straightforward manner. The FAQfinder system is currently available as an internal Web site to the University of Chicago Computer Science community. At present, there are twenty FAQs avsilable to provide answers to user’s questions. Research proceeds in several dixections: cuzrently, a procedure is being developed to automatically separate questions from answers in FAQ files, the number of FAQs available to the user is being expanded, and FAQfinder’s ability to accurately determine the ~im~ex/ty of two natural-language questions is being improved. My own reseaw~ focuses on this final line of research; in this paper, I desex/be an ongoing effort to use a machine lean~g technique (in this case, John Koza’s Genetic Programming) as a means of ;ny~oving FAQfinder’s question-matcJxing performance.